A 17GB model file that beats a closed flagship on the vendor’s own benchmarks is a rare thing, and Qwen 3.8 27B pairs it with a default setting that will quietly waste your whole evening. Released August 15 under an Apache 2 license, the 27B vision-capable model beats its predecessor and even a closed flagship on Qwen’s own benchmarks, yet it ships with reasoning effort set to extra high, which turned a 137-second image generation task into a 21-minute one in Simon Willison’s testing.
The Release
Alibaba’s Qwen lab released Qwen 3.8 27B on Friday, August 15, and the model matters for one structural reason: 27 billion parameters is an excellent size for a laptop. The previous generation, Qwen 3.6 27B, was already one of the better models you could run locally, and the new release is Apache 2 licensed, so it is free to use, modify, and redistribute. The lab’s self-reported benchmarks show a boost over both Qwen 3.6 27B and the closed-weight Qwen 3.7-Plus, which was one of Qwen’s strongest models of any size as recently as May.
Independent testing is still early, but the first serious hands-on review arrived fast. Simon Willison ran the model on two machines, a 128GB M5 Max MacBook Pro and an NVIDIA DGX Spark, using LM Studio and its 17GB Q4_K_M quantized build. A 17GB file for a frontier-adjacent vision model is the whole point of the 27B class: it fits on a reasonably specced laptop, runs entirely offline, and produces results that were the domain of cloud APIs a year ago.
The Default That Wastes Your Time
Here is where it gets weird. Qwen documents the model as defaulting to xhigh for reasoning effort, and the LM Studio GGUF build preserves that default. The model officially supports three reasoning effort settings: xhigh, the default, described as “for complex tasks demanding thorough analysis”; medium, “balancing accuracy and speed”; and low, “efficient reasoning optimizing for speed and cost.”
Willison found the default spectacular, in the worst way. His first test was generating an SVG of a pelican riding a bicycle, a famously hard prompt for image models. With the default xhigh setting, the model spent 21 minutes generating it, using 22,276 reasoning tokens to produce 3,223 tokens of output. The result was genuinely good, he said, the best pelican SVG he has generated with a local model, with a correct bicycle frame, legs on both sides of the bike, and a wing touching the handlebars. But he was blunt about the cost: “Was that worth waiting 21 minutes for? Absolutely not.”
Running the same prompt with reasoning turned off produced a usable result in 137 seconds, just over two minutes, using 3,715 tokens. The output was weaker, the bicycle frame shape was off and the pelican’s feet missed the pedals, but it was still recognizable and took a tenth of the time.
The overthinking is not limited to hard prompts. Willison asked the model to “draw an svg of a circle” and the reasoning trace immediately spiraled into a design project: “The user is asking for an SVG drawing of a circle. Simple request, but I want it to be a carefully crafted piece… a geometric ‘circle study’ with subtle animation, layered rings, and a distinctive palette,” complete with color palette options like “deep teal ink on warm paper” and “Bauhaus/compass-drawing vibe.”
He also hit a practical trap: LM Studio’s default context limit of 8,192 tokens was too small because the model used the entire budget just thinking about mundane problems. Loading the model with the full 262,144 maximum context length fixed that, but it is the kind of configuration detail that will trip up anyone following a quickstart guide.
What the Overthinking Actually Costs
The concrete numbers from Willison’s tests are worth memorizing because they quantify the trade-off:
| Prompt | Reasoning setting | Time | Tokens used |
|---|---|---|---|
| Pelican riding a bicycle (SVG) | xhigh (default) | 21 minutes | 22,276 reasoning + 3,223 output |
| Pelican riding a bicycle (SVG) | Off | 137 seconds | 3,715 total |
| Circle (SVG) | xhigh (default) | minutes of reasoning | entire 8K context budget |
The pattern is consistent: xhigh produces measurably better output on hard creative prompts, but the cost is anywhere from 6x to 10x in time and tokens. For batch work, for agent loops, for anything with a deadline, that default is a trap. The model is excellent, Willison concluded, but “it defaults to wildly overthinking things,” and he called the default “hilarious” in the sense that it is clearly not how anyone should run the model on consumer hardware.
The Bigger Sibling
For completeness, Willison also ran the same pelican prompt through the much larger Qwen 3.8 2.4T-A95B, a MoE model released the previous week, via OpenRouter, and got an animated SVG. That is a reminder that the 27B release is one point on a spectrum: if you have cloud budget or a big server, the 2.4T parameter MoE is the ceiling of the family, while 27B is the floor that runs on a laptop. For local deployment, 27B is the interesting one, and the reason is the same size argument that has made other sub-30B models popular for on-device work. We compared Qwen 3.8’s general release earlier this year in our Qwen 3.8 caught up to Claude piece, and if you are shopping local reasoning models, our MiniMax H3 MLX on Apple Silicon review covers the other strong local option and how it handles the same trade-offs.
The Honest Caveats
The benchmarks are Qwen’s own until independent labs weigh in, and the “beats Qwen 3.7-Plus” claim deserves a skeptical eye given that 3.7-Plus is a much larger, presumably more expensive model. Willison’s tests are a single user’s experience on two machines, and quantized builds can behave differently from the full model. The xhigh default is documented, so it is not a bug, but it is a product decision that penalizes anyone who does not read the docs, and most users will not.
There is also the context-budget interaction. Even at 262K context, a model that burns 22K reasoning tokens on a single SVG is going to exhaust the window fast on real tasks, which means more time waiting and more careful prompt design, not less. Teams planning to run this model in production should treat reasoning effort as a hyperparameter to tune per workload, not a global setting.
Who Should Care
Local-first developers get the best deal here: a 17GB model that handles vision, coding, and creative generation offline. Anyone running agent loops locally should set reasoning effort to medium or low and reserve xhigh for single-shot creative tasks where quality is the whole point. LM Studio users should bump the context window before doing anything else. And anyone who was waiting for a laptop-sized model with genuine vision capability has their answer, provided they are willing to change one default setting first.
The Bottom Line
Qwen 3.8 27B is an excellent model with a strange default. The capability is real, the 17GB footprint is the category’s sweet spot, and the benchmark story is strong, but the xhigh reasoning default makes the model feel ten times slower than it needs to be for routine work. The fix is one setting change. Once you make it, you have a genuinely impressive local vision model that runs on a laptop and costs nothing per token. That is a rare combination, and worth the minute it takes to configure.